Using mean-squared error to assess visual image quality

Conclusions about the usefulness of mean-squared error for predicting visual image quality are presented in this paper. A standard imaging model was employed that consisted of: an object, point spread function, and noise. Deconvolved reconstructions were recovered from blurred and noisy measurements formed using this model. Additionally, image reconstructions were regularized by classical Fourier-domain filters. These post-processing steps generated the basic components of mean-squared error: bias and pixel-by-pixel noise variances. Several Fourier domain regularization filters were employed so that a broad range of bias/variance tradeoffs could be analyzed. Results given in this paper show that mean-squared error is a reliable indicator of visual image quality only when the images being compared have approximately equal bias/variance ratios.